Scribd
Upload
56 views15 pages

Attn Is All You Need

Attention is all you need by google

Uploaded by

anveshnayak17
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
56 views15 pages

Attn Is All You Need

Attention is all you need by google

Uploaded by

anveshnayak17
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
You are on page 1/ 15
arXiv:1706.03762v7 [cs.CL] 2 Aug 2023 Provided proper attribution is provided, Google hereby grants permission to reproduce the tables and figures in this paper solely for use in journalistic or scholarly works. Attention Is All You Need [Ashish Vaswani* Noam Shazeer* Niki Parmar” Jakob Uszkoreit” Google Brain Google Brain Google Research Google Research [email protected] [email protected] [email protected] [email protected], Lion Jones Aidan N. Gomez" | Lukasz Kaiser" Google Research, University of Toronto Google Brain Lliontgoogle.con [email protected] [email protected] Ilia Polosukhin” * illia.polosukhintgmail com Abstract ‘The dominant sequence transduction models are based on complex recurrent or convolutional neural networks that include an encoder and a decoder. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, ‘based solely on attention mechanisms, dispensing with recurrence and convolutions ‘entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English- to-German translation task, improving over the existing best results, including ‘ensembles, by over 2 BLEU. On the WMT 2014 English-to-French translation task, ‘our model establishes a new single-model state-of-the-art BLEU score of 41,8 after taining for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature, We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data “Enqual contibation, Lasting order i random. Jakob proposed replacing RNNs with sef-attention and started. the effort to evaluate tis idea. Ashish, with Ilia. designed and implemented th frst Transformer models and has been crucially involved in every aspect of thie work. Noam proposed scaled do-product attention, ntuli-head attention andthe parameter-ree postion representation and became the other person involved in neatly every ‘detail, Niki designed, implemented, taned and evaluated countless model variants in our original codebase and. tensor2tensor. Lhon also experimented with novel mode! vaians, Was responsible for our initial codebase, and ‘efficient inference and visualizations. Lukasz and Aidan spent countless long days designing Various parts of and implementing tensor2tensor, replacing ou: eatier codebase, greatly improving results and massively accelerating ‘our research, ‘Work performed while at Google Bran, + Work performed while at Google Research, 1 Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA. 1 Introduction Recurrent neural networks, ong short-term memory [13] and gated recurrent [7] neural networks in particular, have been firmly established as state of the art approaches in sequence modeling and. transduction problems such as language modeling and machine translation (35, 2, 5]. Numerous. efforts have since continued to push the boundaries of recurrent language models and encoder-decoder, architectures [38, 24, 15] Recurrent models typically factor computation along the symbol positions of the input and output sequences. Aligning the positions to steps in computation time, they generate a sequence of hidden states Jy, a8 a function ofthe previous hidden state hy and the input for position ¢. This inherently sequential nature precludes parallelization within training examples, which becomes critical at longer sequence lengths, as memory constraints limit batching across examples. Recent work has achieved significant improvements in computational efficiency through factorization tricks [21] and conditional ‘computation [32], while also improving model performance in case of the later. The fundamental ‘constraint of sequential computation, however, remains. Attention mechanisms have become an integral part of compelling sequence modeling and transduc- tion models in various tasks, allowing modeling of dependencies without regard to their distance in the input or output sequences [2, 19]. In all buta few cases [27], however, such attention mechanisms are used in conjunction with a recurrent network. In this work we propose the Transformer, a model architecture eschewing recurrence and instead. relying entirely on an attention mechanism to draw global dependencies between input and output ‘The Transformer allows for significantly more parallelization and can reach a new state of the art in translation quality after being trained for as litte as twelve hours on eight P100 GPUs. 2 Background ‘The goal of reducing sequential computation also forms the foundation ofthe Extended Neural GPU {16}, ByteNet [18] and ConvS2S (9), al of which use convolutional neural networks as basic building block, computing hidden representations in parallel forall input and ouput positions. In these models, the number of operation required to relate signal from two arbitrary input or output positions grows in the distance between positions, linearly for ConvS2S and logarithmically for ByteNet, This makes it more difficult to learn dependencies between distant positions [12]. In the Transformer this is reduced to a constant number of operations, albeit atthe cost of reduced effective resolution due to averaging attention-Weighted positions, an effect we counteract with Mul-Head Attention as described in section 3.2. ‘Self-atention, sometimes called intra-attention is an attention mechanism relating different positions ‘ofa single sequence in order to compute a representation of the sequence. Self-attention has been. ‘used successfully in a variety of tasks including reading comprehension, abstractive summarization, textual entailment and learning task-independent sentence representations (4, 27, 28, 22] End-to-end memory networks are based on a recurrent attention mechanism instead of sequence- aligned recurrence and have been shown to perform well on simple-language question answering and language modeling tasks [34] ‘To the best of our knowledge, however, the Transformer is the first transduction model relying centiely on self-attention to compute representations of its input and output without using sequence- ‘aligned RNNs or convolution. In the following sections, we will describe the Transformer, motivate self-attention and discuss its advantages over models such as [17, 18] and [9] 3 Model Architecture “Most competitive neural sequence transduction models have an encoder-decoder structure [5, 2, 35] Here, the encoder maps an input sequence of symbol representations (r1,..z_) t0 a sequence ‘of continuous representations % = (21,..,2n)- Given 2, the decoder then generates an output sequence (9. 3x) of symbols one element ata time. At each step the model is auto-regressive [10}, consuming the previously generated symbols as additional input when generating the next Output Probabiltes cam 4 boss: Treat usa Embedding Emeedding Inputs Outputs (ehitted right Figure 1: The Transformer - model architecture, ‘The Transformer follows this overall architecture using stacked self-attention and point-wise, fully ‘connected layers for both the encoder and decoder, shown in the left and right halves of Figure 1, respectively, 3.1 Encoder and Decoder Stacks Encoder: The encoder is composed of a stack of N= 6 identical layers. Each layer has two sub-layers, The first is a multi-head self-attention mechanism, and the second isa simple, position ‘wise fully connected feed-forward network. We employ a residual connection [11] around each of the two sub-layers, followed by layer normalization [1]. That is, the output of each sub-layer is LayerNorm(sr + Sublayer(2.)), where Sublayer() is the function implemented by the sub-layer itself. To facilitate these residual connections, all sub-layers in the model, as well as the embedding layers, produce outputs of dimension doéai = 512. Decoder: The decoder is also composed of a stack of V ~ 6 identical layers. In addition tothe wo sub-layers in each encoder layer, the decoder inserts a third sub-layer, which performs multi-head attention over the output of the encoder stack. Similar to the encoder, we employ residual connections around each of the sub-layers, followed by layer normalization. We also modify the self-attention, sub-layer in the decoder stack to prevent positions from attending to subsequent positions. This ‘masking, combined with fat thatthe output embeddings are offset by one position, ensures that the predictions for position ican depend only on the known outputs at positions less than 1 3.2 Attention ‘An atlention function can be described as mapping a query and a set of key-value pairs to an output, ‘where the query, keys, values, and output are all vectors. The output is computed as a weighted sum Scaled Dot-Product Attention “Multi-Head Attention t Ther Math Ganesh (otnex Mask (58), Beate Cam) near) (Linear) (near ~~) aK v vk @ Figure 2: (left) Scaled Dot-Product Attention. (right) Multi-Head Attention consists of several attention layers running in parallel, of the values, where the weight assigned to each value is computed by a compatibility function of the ‘query with the corresponding key. 3.2.1 Sealed Dot-Product Attention We call our particular attention "Scaled Dot-Product Attention" (Figure 2). The input consists of «queries and Keys of dimension dl,, and values of dimension d. We compute the dot products of the ‘query with all keys, divide each by /e, and apply a softmax function to obtain the weights on the values, In practice, we compute the attention function on a set of queries simultancously, packed together into a matrix Q. The keys and values are also packed together into matrices A and V. We compute the matrix of outputs as Qk? AXtention(@, KV) = softmax( 4 wo ‘The two most commonly used attention functions are additive attention [2], and dot-product (multi- plicative) attention. Dot product attention is identical to our algorithm, except forthe sealing factor of aie. Additive ateation computes the compatiblity function using a Leed-forwazd network with a single bidden layer, While the two are similar in theoretical complenity, do-produet attention is uch faster and more space-efficient in practice, since it ean be implemented using highly optimized matrix maltipication code While for small values of dy the two mechanisms perform similarly additive attention outperforms dot product attention without scaling for larger values of dy [3]. We suspect that for large values of ‘ty, the dot products grow large in magnitude, pushing the softmax function into regions where it has extremely small gradients‘. To counteract this elect, we sale the dot products by 3.2.2 Multi-Head Attention Instead of performing a single attention function with dgge-dimensional keys, values and queries, we found it beneficial to linearly project the queries, Keys and values h times with different, learned linear projections to di, dy and dy dimensions, respectively. On each of these projected versions of queries, keys and values we then perform the attention function in parallel, yielding d, dimensional “Po llsrate why he dot produce ge large, assume tha the componcas of and kare independent random variables with mean O and vatiance 1. Then ther dot product, q- = So, qk, bas mean O ad variance di ‘output values, These are concatenated and once again projected, resulting in the final values, as depicted in Figure 2. ‘Multi-head attention allows the model to jointly attend to information from different representation subspaces at different positions. With a single attention head, averaging inhibits this. Multiffead(@, KV) = Concat(headh,.., head,)1V° contion(QW 2, KW, VY) where head = Where the projections are parameter matrices W2 € Rt, WK © Rents, WY © Reeirte and WO e Rito In this work we employ h — 8 parallel attention layers, or heads, For each of these we use dy = dy = dyosa/t = G4. Due to the reduced dimension of each head, the total computational cost is similar to that of single-head attention with full dimensionality. 3.2.3 Applications of Attention in our Model ‘The Transformer uses multi-head attention in three different ways: + In “encoder-decoder attention” layers, the queries come from the previous decoder layer, ‘and the memory keys and values come from the output of the encoder. This allows every position in the decoder to attend over all positions in the input sequence. This mimics the typical encoder-decoder attention mechanisms in sequence-Co-sequence models such as (38. 2,9). ‘The encoder contains self-attention layers. Ina self-attention layer all of the keys. values and queries come from the same place, in this case, the output of the previous layer in the ‘encoder. Each position in the encoder can attend to all positions in the previous layer of the encoder, ‘Similarly, seU-attention layers in the decoder allow each position in the decoder to attend to all positions in the decoder up to and including that position. We need to prevent leftward information flow in the decoder to preserve the auto-regressive property. We implement this inside of scaled dot-product attention by masking out (setting to —>c) all values in the input of the softmax which correspond to illegal connections. See Figure 2 33 Position-wise Feed-Forward Networks In addition to attention sub-layers, each of the layers in our encoder and decoder contains a fully ‘connected feed-forward network, which is applied to each position separately and identically. This ‘consists of two linear transformations with a ReLU activation in between. FFN(2) = max(0, 20; +5:)Wa +b @ While the linear wansformations are the same across different positions, they use different parameters from layer to layer. Another way of describing this is as two convolutions with kemel size 1 The dimensionality of input and output is dygas) = 512, and the inner-layer has dimensionality dys = 2088. 3.4 Embeddings and Softmax Similarly to other sequence transduction models, we use learned embeddings to convert the input tokens and output tokens to vectors of dimension dang. We also use the usual learned linear transfor ‘mation and softmax function to convert the decoder output to predicted next-token probabilities. In ‘our model, we share the same weight matrix between the two embedding layers and the pre-softmax linear transformation, similar to [30]. In the embedding layers, we multiply those weights by Vm. Table 1: Maximum path lengths, per-layer complexity and minimum number of sequential operations for different layer types. n is the sequence length, dis the representation dimension, is the kernel size of convolutions and r the size of the neighborhood in restricted self-atention. Layer Type ‘Complexity per Layer Sequential Maximum Path Length Operations Saran OC SCYSC“‘(‘COCCOCOMSCOOOCOC™ Recurrent O(n) O(n) O(n) Convolutional O(k-n-d) 0) Oltogs(n)) Self-Attention (restricted) Olr-n-a ou) O(njr) 34 Positional Encoding Since our model contains no recurrence and no convolution, in order for the model to make use of the ‘order of the sequence, we must inject some information about the relative or absolute position ofthe tokens in the sequence. To this end, we add "positional encodings" to the input embeddings at the bottoms of the encoder and decoder stacks. The positional encodings have the same dimension dace as the embeddings, so that the two can be summed, There are many choices of positional encodings, eared and fixed [9] In this work, we use sine and cosine functions of different frequencies: PEjgo. = sin(pos/10000%/*) PEhpos,ais1) = €08(pos/10000%/H-#') ‘where pos is the postion and é is the dimension, That is, each dimension of the positional encoding corresponds toa sinusoid, The wavelengths form a geometric progression from 27 to 10000 2x, We chose this function because we hypothesized it would allow the model to easly learn to attend by relative positions, since for any fixed offset k, PEpos 4: can be represented as a linear function of PE, We also experimented with using learned positional embeddings [9] instead, and found thatthe two versions produced nearly identical results (see Table 3 row (E)). We chose the sinusoidal version because it may allow the model to extrapolate to sequence lengths longer than the ones encountered during training 4) Why Self-Attention In this section we compare various aspects of self-atention layers to the recurrent and convolu- tional layers commonly used for mapping one variable-length sequence of symbol representations (21,24) to another sequence of equal length “eu). with 2,,2; € RS, such as a hidden layer in @ typical sequence transduction encoder or decoder. Motivating our use of sel-attention we consider three desiderata ‘One is the total computational complexity per layer. Another is the amount of computation that can be parallelized, as measured by the minimum number of sequential operations required. ‘The third is the path length between long-range dependencies in the network, Learning long-range dependencies isa key challenge in many sequence transduction tasks. One key factor affecting the ability to leam such dependencies is the length of the paths forward and backward signals have to iraverse in the network. The shorter these paths between any combination of positions in the input ‘and output sequences, the easier itis to earn long-range dependencies [12]. Hence we also compare ‘the maximum path length between any two input and output positions in networks composed of the different layer types. As noted in Table 1, a self-atention layer connects all positions with a constant number of sequentially ‘executed operations, whereas a recurrent layer requires O(n) sequential operations, In terms of ‘computational complexity, self-attention layers are faster than recurrent layers when the sequence length n is smaller than the representation dimensionality d, which is most often the case with sentence representations used by state-of-the-art models in machine translations, such as word-piece [38] and byte-pair [31] representations. To improve computational performance for tasks involving very long sequences, sel-altention could be restricted to considering only a neighborhood of size r in the input sequence centered around the respective output position, This would inerease the maximum. path length to O(n). We plan to investigate this approach further in future work, A single convolutional layer with kernel width &: heeeigeggs 2993988 le 2 i 3 BGRRERE & ae = = Figure 3: An example of the attention mechanism following long-distance dependencies in the ‘encoder self-attention in layer 5 of 6, Many of the attention heads attend to a distant dependency of the verb ‘making’, completing the phrase ‘making...more difficult’. Attentions here shown only for the word ‘making’. Different colors represent different heads. Best viewed in color. : Uuojuido vuotuido fw fu uw 4 Buss Suissi awe ue om om yeu eu sy si asnt asnt °q oq pinoys pinoys uuoneaiidde. sy nq weeped eq eu, oreo GE Figure 4: Two attention heads, also in layer 5 of 6, apparently involved in anaphora resolution. Top: Full atentions for head 5. Bottom: Isolated attentions {rom just the word ‘its’ for atention heads 5 and 6. Note that the attentions are very sharp for this word. 4 pinoys“ uuoneoidde pinoys dae ——___—

You might also like